Update models.py

models.py

@@ -4,6 +4,7 @@ import numpy as np
 
 classifier = None
 
+
 def load_model():
     global classifier
     if classifier is None:
@@ -20,41 +21,59 @@ def detect_audio(y: np.ndarray) -> tuple[str, float, str]:
     Detect if audio is AI_GENERATED or HUMAN.
     Returns: classification, confidenceScore (0-1), explanation
     """
+
     try:
-        result = load_model()
+        # ✅ Always ensure model is loaded
+        model = load_model()
+
+        # ✅ Pass correct input format
+        result = model(
+            {
+                "array": y,
+                "sampling_rate": 16000
+            }
+        )
+
         if not result:
             return "HUMAN", 0.50, "Insufficient audio features detected."
 
         # Take top prediction
         top = result[0]
-        label_lower = top['label'].lower()
-        top_score = top['score']
+        label_lower = top["label"].lower()
+        top_score = top["score"]
 
-        # Flexible mapping for common labels
-        if any(word in label_lower for word in ['ai', 'fake', 'synthetic', 'aivoice']):
+        # Label mapping
+        if any(word in label_lower for word in ["ai", "fake", "synthetic", "aivoice"]):
             classification = "AI_GENERATED"
-            confidence = round(top_score, 3)
         else:
             classification = "HUMAN"
-            confidence = round(top_score, 3)
 
-        # Feature-based explanation (judge-friendly)
+        confidence = round(float(top_score), 3)
+
+        # Feature analysis
         flatness = librosa.feature.spectral_flatness(y=y).mean()
+
         pitch = librosa.yin(y, fmin=75, fmax=300)
         pitch_std = np.std(pitch) if len(pitch) > 0 else 0.0
 
         cues = []
+
         if flatness > 0.5:
             cues.append("unnatural high spectral flatness (robotic)")
         else:
             cues.append("natural spectral variation")
+
         if pitch_std < 10:
             cues.append("unnatural pitch consistency")
         else:
             cues.append("natural pitch variation")
 
-        # Decide feature-based tendency
-        feature_vote = "AI_GENERATED" if (flatness > 0.5 and pitch_std < 10) else "HUMAN"
+        # Feature vote
+        feature_vote = (
+            "AI_GENERATED"
+            if (flatness > 0.5 and pitch_std < 10)
+            else "HUMAN"
+        )
 
         cues_text = " and ".join(cues)
 
@@ -74,5 +93,9 @@ def detect_audio(y: np.ndarray) -> tuple[str, float, str]:
         return classification, confidence, explanation
 
     except Exception as e:
-        # Fallback on error
-        return "HUMAN", 0.50, f"Analysis error: {str(e)}. Treated as human."
+
+        return (
+            "HUMAN",
+            0.50,
+            f"Analysis error: {str(e)}. Treated as human."
+        )